AlphaPulldown + AlphaFold3 Research Log

This README documents the complete pipeline, setup, troubleshooting, and experimental execution history for running AlphaPulldown and AlphaFold3 workflows for protein complex structure prediction. All content is designed for reproducibility and future reference.

Table of Contents

• TODO
• Reference Links
• File Transfer & Utilities
• Expected Data from AlphaFold3
• AlphaPulldown Overview
• Environment Setup
• Feature Generation
• Understanding create_individual_features.py
• Accessing Databases
• Multimer Job Execution
• Result Analysis
• AlphaFold3 Notes
• Troubleshooting & Exceptions

---

TODO

1. Check missing_proteins.txt for Feature Database completeness.
2. Investigate MMseqs2 errors and output inconsistencies.

---

Reference Links

UniProt

• Protein sequences: https://www.uniprot.org
• REST API example: https://www.uniprot.org/uniprotkb/Q04740/entry

AlphaPulldown

• Paper: https://academic.oup.com/bioinformatics/article/39/1/btac749/6839971
• GitHub: https://github.com/KosinskiLab/AlphaPulldown

AF2-Multimer

• SBGrid example: https://sbgrid.org/wiki/examples/alphafold2
• GitHub: https://github.com/google-deepmind/alphafold

AlphaFold3

• Server: https://alphafoldserver.com/welcome
• GitHub: https://github.com/google-deepmind/alphafold3

---

File Transfer & Utilities

# Transfer files (adjust paths as needed):
scp /local/path/to/file jzho349@kilimanjaro.biochem.emory.edu:/remote/path
scp -r jzho349@kilimanjaro.biochem.emory.edu:/remote/result/path ~/Downloads/

# Split protein FASTA list into 8 chunks:
split -n l/8 --suffix-length=2 --additional-suffix=.txt unique_protein_ids_Jack.txt target_chunk_Jack_

# One-liner tracing number of jobs completed
for d in gpu_*; do echo "$d: $(ls -1 "$d" | wc -l)"; done

---

Expected Data from AlphaFold3

1. mmCIF: Predicted complex structure.
2. pLDDT: Per-residue confidence score.
3. Possibly more outputs such as ipTM or pairwise confidence (to be explored).

---

AlphaPulldown Overview

AlphaPulldown consists of two core stages:

1. create_individual_features.py:

   • Computes MSAs and finds templates.
   • Stores monomer features as .pkl files.

2. run_multimer_jobs.py:

   • Predicts structures based on the generated features.

---

Environment Setup

# Create environment
conda create -n AlphaPulldown -c omnia -c bioconda -c conda-forge \
  python==3.11 openmm==8.0 pdbfixer==1.9 kalign2 hhsuite hmmer modelcif

# Activate & install JAX for GPU acceleration
conda activate AlphaPulldown
pip install -U "jax[cuda12]"

---

Feature Generation

MMseqs2 (Sequential Mode)

NUM_SEQ=$(grep -c "^>" protein_sequences.fasta)
for ((i=0; i<$NUM_SEQ; i++)); do
  echo "Processing index $i"
  create_individual_features.py \
    --fasta_paths=$FASTA_PATH \
    --data_dir=$DATA_DIR \
    --output_dir=$OUTPUT_DIR \
    --skip_existing=True \
    --use_mmseqs2=True \
    --max_template_date="2050-01-01" \
    --seq_index=$i || echo $i >> failed_indexes.txt

done

JackHMMer (Parallelized)

MAX_JOBS=16
rm -f $FAILED_LOG

for ((i=0; i<$NUM_SEQ; i++)); do
  (
    create_individual_features.py \
      --fasta_paths=$FASTA_PATH \
      --data_dir=$DATA_DIR \
      --output_dir=$OUTPUT_DIR \
      --skip_existing=True \
      --use_mmseqs2=False \
      --max_template_date="2050-01-01" \
      --seq_index=$i || echo $i >> $FAILED_LOG
  ) &

  while (( $(jobs -r | wc -l) >= MAX_JOBS )); do sleep 1; done

done
wait

---

Understanding create_individual_features.py

• CPU-only script
• Output format: Pickle files (.pkl)
• Requires pre-downloaded Genetic Database
• Feature sources:
  1. JackHMMer/HHblits (default)
  2. MMseqs2
  3. Precomputed Feature Pickles

Resources

• Genetic DB: https://github.com/KosinskiLab/alphafold#genetic-databases
• Feature DB: https://alphapulldown.s3.embl.de/index.html

---

Accessing Databases

Genetic Database

bash download_all_data.sh /data7/Conny/data/AF_GeneticDB

Feature Pickle Database

# Install MinIO
curl -O https://dl.min.io/client/mc/release/linux-amd64/mc
chmod +x mc && mkdir -p $HOME/bin && mv mc $HOME/bin/
echo 'export PATH=$HOME/bin:$PATH' >> ~/.bashrc && source ~/.bashrc

# Example download
mc cp embl/alphapulldown/input_features/Saccharomyces_cerevisiae/Q01329.pkl.xz /data7/Conny/data/JackFeaturePickleDB

# Batch download
bash download_found.sh | tee download_output.log
xz -dk *.xz  # decompress while keeping original

---

Multimer Job Execution

Parallel execution across 8 GPUs:

TARGET_FILES=(
  /data7/Conny/data/target_chunk_aa
  /data7/Conny/data/target_chunk_ab
  ...
  /data7/Conny/data/target_chunk_ah
)

for i in ${!TARGET_FILES[@]}; do
  CUDA_VISIBLE_DEVICES=$i run_multimer_jobs.py \
    --mode=pulldown \
    --monomer_objects_dir=$MONOMER_OBJECTS_DIR \
    --protein_lists=$(realpath $BAIT_FILE),${TARGET_FILES[$i]} \
    --output_path=$BASE_OUTPUT/gpu_$i \
    --data_dir=$DATA_DIR \
    --num_cycle=3 \
    --num_predictions_per_model=1 \
    > $BASE_OUTPUT/gpu_$i/run.log 2>&1 &
done

wait

---

Result Analysis

1. Create Jupyter notebooks using:

create_notebook.py --cutoff=5.0 --output_dir=/path/to/results

2. Visualize outputs (e.g., PAE, ipTM, ranked pdbs)
3. For full table generation:
   • Build apptainer image with CCP4 libraries
   • Execute:

apptainer exec --no-home --bind /results:/mnt fold_analysis_final.sif /app/run_get_good_pae.sh --output_dir=/mnt --cutoff=10

---

AlphaFold3 Notes

• GUI: Automated MSA
• Local mode: Accepts MSA input via JSON
• Additional work needed to automate Selenium upload process (autoclicker test in progress)

---

Troubleshooting & Exceptions

1. UniRef30 issue → Fix with: https://github.com/google-deepmind/alphafold/pull/860/files
2. MMseqs2: 5 failures → P40449, P23369, Q12754, Q22354, Q12019
3. JackHMMer: 3 known failures → P40328, P48570, P32432，P38691，P48510，P53920，Q00582 → cd /data7/Conny/result_JackHMMer/missing_pkls.txt
4. FeatureDB: 20 missing pickles → cd /data7/Conny/data/JackFeaturePickleDB/missing_pickle_proteins.txt
5. AF3: 1 missing pickles due to large size → Job_999

---

This document is actively maintained as experimental runs progress.